Case study · Data product, designed and built

Five seasons of Formula 1, turned into a site you can question.

RI360 takes raw race timing data and makes it readable: who was fast, what each tyre cost, and what would have happened on a different strategy. I built the data pipeline, designed the screens and wrote the site.

Project
Self-initiated product
My role
Data pipeline, design, front end
Built with
Python, FastAPI, React, TypeScript
Status
Live
  • 96,223laps analysed
  • 89races, 2019 to 2023
  • 86race models fitted
  • 100%of checked lap times match the raw feed
the product

One design system, two themes

Paper is for reading an analysis. Pit wall is for watching numbers. Both come from the same set of design tokens, so every page works in either. Drag the handle to switch.

RI360 overview page in the light paper theme: the headline 96,223 laps, one pit wall, beside a race trace chart.
The same RI360 overview page in the dark pit wall theme.
PaperPit wall
Drag to switch theme · live site, Nov 2026
the problem

The data was counting laps twice

Before designing anything I profiled the lap table the first version of this project had exported. Almost a quarter of it was wrong.

  1. 124,209

    rows in the exported lap table

    This is what the earlier model had been trained on.

    100% of the original export
  2. 27,986

    of them were the same lap stored twice

    23,038 in 2022 and 4,948 in 2023. That is 22.5% of the export. The duplicates had also doubled the stint-age value the old model learned from.

    22.5% of the original export
  3. 96,223

    real laps left after removing them

    I also rebuilt 612 missing tyre ages and kept 368 laps with an unknown tyre out of the tyre analysis.

    77.5% of the original export
  4. 94,490

    lap times checked against the raw timing feed

    Every one of them matches to the millisecond. Only then did I build charts on top.

    76.1% of the original export
the build

Five decisions and the reason for each

Each one traded something away. I have said what.

  1. 01

    One equation instead of a black box

    Problem
    The first version used a Random Forest. It could not explain a prediction, and it had learned from the duplicated laps.
    What I did
    Each race now gets one fitted equation: lap time = driver + fuel × laps left + tyre compound + wear × tyre age. Its median error on laps held out from training is 0.47 seconds, across 86 races.
    What it costs
    It treats tyre wear as a straight line, and three wet races get no model at all.
  2. 02

    Measure the pit stop, don't assume it

    Problem
    The old simulator charged a fixed 22 seconds for every pit stop at every circuit.
    What I did
    I measure it from the laps: in-lap plus out-lap, minus two normal laps. A stop costs about 18.6 seconds at Spa and 29.5 at Lusail, which decides whether an extra stop can ever pay.
    What it costs
    It needs enough green-flag stops at a circuit to be trusted.
  3. 03

    Grade every model in public

    Problem
    A model that fits one race well can fit another badly, and a chart hides that.
    What I did
    Every fit carries a grade. 60 races have a wear rate per tyre, 12 share one rate, 7 are marked weak and 7 are mixed wet-and-dry races. The Method page lists what the model does not know.
    What it costs
    Some races show less than others, on purpose.
  4. 04

    No server

    Problem
    A site that needs a database and an API is slow to load and costs money to keep running.
    What I did
    The pipeline writes everything as 4.6 MB of static files: one index and one file per race. The strategy lab runs the equation in the browser and tries every one-stop and two-stop plan.
    What it costs
    New races need the pipeline to be run again and the site redeployed.
  5. 05

    Draw the charts by hand

    Problem
    Chart libraries bring their own look, and a race trace with twenty drivers needs careful control.
    What I did
    Every chart is hand-built SVG, using d3 only for scales and line shapes, styled from the same tokens as the rest of the site.
    What it costs
    Each new chart type is real work.
try it

Run the model yourself

This is the real fitted equation for the 2022 Abu Dhabi Grand Prix, running in your browser with the numbers from the site's data file. Pick the tyres and the pit lap.

2022 Abu Dhabi Grand Prix · 58 laps · winner's pace

First stint
Second stint

Medium 20 laps, Hard 38 laps

Predicted race time
1:27:44.6
Compared with the fastest one-stop plan
+0.0s

This is the fastest one-stop plan the model can find. It is also the plan the winner really ran: medium tyres for 20 laps, then hard.

No traffic, safety cars or tyre warm-up. Pit stop cost measured for this race: 24.0 seconds.

the screens

Six pages, each with one job

Choose a page to see it.

Headline findings from the data: what a pit stop costs at each circuit, how fast each tyre wears, and who dominated each season.

RI360 findings: a bar chart of pit stop cost by circuit and tyre wear rates for soft, medium and hard.

On a phone

RI360 overview on a phone in the paper theme.
Overview · paper
RI360 race page on a phone in the pit wall theme.
Race · pit wall
the design file

Designed in Figma before it was built

The Figma file holds every page in three versions, plus the shared parts.

  • 6pages designed for desktop in the paper theme
  • 6of the same pages for a 390 px phone
  • 6again for desktop in the pit wall theme
  • 3shared components: the site header and two footers

Type is Barlow Condensed for headlines, Space Grotesk for text and JetBrains Mono for numbers.

Open the Figma file
under the hood

From timing feed to website in four steps

  1. 01

    Collect

    A Python script uses the open-source FastF1 library to download race timing, tyre and weather data. A FastAPI service stores it in SQL: SQLite on my machine, Postgres in production.

  2. 02

    Clean

    Remove duplicate laps, rebuild tyre ages, and compare every lap against the raw timing feed. Safety-car and red-flag periods are mapped to laps from the track-status feed.

  3. 03

    Model

    Fit one pace equation per race on clean laps only, and score it on a random 20% of laps it has not seen.

  4. 04

    Ship

    Write the result as static files and build the site. GitHub Actions checks on every push that the site builds and the API starts.

limits

What it does not do

  • The model knows nothing about traffic, dirty air, safety-car timing or tyre warm-up. It compares plans; it does not predict races.
  • The 2021 season holds rounds 1 to 7 only.
  • The live site does not use the API. All its data ships with it.

Timing data comes from the FastF1 library. RI360 is an independent project and is not affiliated with Formula 1, the FIA or any team.

Have data that nobody can read yet?

I can clean it, model it and design the product that shows it.